OpenAI Whisper Download: Free Speech-to-Text You Run Locally
Whisper 20250625 · for Windows, macOS and Linux
Developer: OpenAI (opens in a new tab)
- Windows
- macOS
- Linux
The download link points to the program’s official source; the file is not hosted on this site. Link address: github.com
Link not working? Let us know
Whisper is OpenAI’s open-source speech recognition model: it transcribes speech in many languages, translates it into English and identifies the spoken language, all on your own computer. The latest release is 20250625, published on June 26, 2025; it installs with pip on Windows, macOS and Linux and is MIT-licensed.
Whisper is the model behind a large share of today’s automatic subtitles and transcripts. OpenAI released both the code and the model weights under the MIT license, so you can run it on your own computer without sending audio anywhere. The repository has about 112,000 stars on GitHub as of October 2026.
What does Whisper do?
- Transcription: Converts speech in audio and video files into text in many languages.
- Translation: Translates speech from other languages into English.
- Language detection: Identifies which language is being spoken.
- Several model sizes: From the tiny model to large and turbo, so you can trade accuracy for speed. English-only .en versions of the smaller models tend to be more accurate for English.
Who uses it?
It is useful for journalists, students, podcasters and developers who need accurate transcripts and want to keep their recordings private.
Installing Whisper and ffmpeg
Whisper is a Python package. Install or update it with:
pip install -U openai-whisper
It also needs the ffmpeg command-line tool:
# Ubuntu or Debian
sudo apt update && sudo apt install ffmpeg
# macOS (Homebrew)
brew install ffmpeg
# Windows (Chocolatey)
choco install ffmpeg
If pip fails while building tiktoken, the README suggests installing Rust, because a prebuilt wheel may not exist for your platform.
First steps
Transcribe a file with the turbo model:
whisper audio.mp3 --model turbo
To translate Japanese speech into English, use a multilingual model (turbo is not trained for translation):
whisper japanese.wav --model medium --language Japanese --task translate
Model sizes and VRAM
| Model | Parameters | VRAM | Relative speed |
|---|---|---|---|
| tiny | 39 M | ~1 GB | ~10x |
| base | 74 M | ~1 GB | ~7x |
| small | 244 M | ~2 GB | ~4x |
| medium | 769 M | ~5 GB | ~2x |
| large | 1550 M | ~10 GB | 1x |
| turbo | 809 M | ~6 GB | ~8x |
Speeds are relative to the large model, measured on an NVIDIA A100. Whisper was trained and tested with Python 3.9.9 and is expected to work with Python 3.8 to 3.11 and recent PyTorch versions.
License
Whisper’s code and model weights are released under the MIT license.
Pros
- Accurate transcription in many languages
- Runs locally, so recordings stay private
- Model sizes for slow and fast hardware
Cons
- Command-line only, no graphical interface
- Large models need a powerful GPU
Frequently Asked Questions
Which Whisper model should I use?
Turbo is a good default: it is about eight times faster than large and needs around 6 GB of VRAM. For translation into English use a multilingual model such as medium or large, because turbo is not trained for translation.
Does Whisper need ffmpeg?
Yes. Whisper uses ffmpeg to read audio files. Install it with your system’s package manager, such as apt, Homebrew, Chocolatey or Scoop.
WhisperOpen source
Download (opens in a new tab)
